[FFmpeg-devel] [PATCH v2 2/5] avutil/mips: Extract load store with shift C1 pair marco
yinshiyou-hf at loongson.cn
yinshiyou-hf at loongson.cn
Thu Jul 22 14:55:20 EEST 2021
> -----原始邮件-----
> 发件人: "Jiaxun Yang" <jiaxun.yang at flygoat.com>
> 发送时间: 2021-07-21 17:19:10 (星期三)
> 收件人: ffmpeg-devel at ffmpeg.org
> 抄送: yinshiyou-hf at loongson.cn, "Jiaxun Yang" <jiaxun.yang at flygoat.com>
> 主题: [FFmpeg-devel] [PATCH v2 2/5] avutil/mips: Extract load store with shift C1 pair marco
>
> We're doing some fancy hacks with load store with shift C1
> beside unaligned load store. Create a marco for l/r pair
> to allow us use it in these places.
>
> Signed-off-by: Jiaxun Yang <jiaxun.yang at flygoat.com>
> ---
> libavutil/mips/mmiutils.h | 49 ++++++++++++++++++++++++---------------
> 1 file changed, 30 insertions(+), 19 deletions(-)
>
> diff --git a/libavutil/mips/mmiutils.h b/libavutil/mips/mmiutils.h
> index 41715c6490..f5b600e50c 100644
> --- a/libavutil/mips/mmiutils.h
> +++ b/libavutil/mips/mmiutils.h
> @@ -57,8 +57,9 @@
> #define MMI_LWC1(fp, addr, bias) \
> "lwc1 "#fp", "#bias"("#addr") \n\t"
>
> -#define MMI_ULWC1(fp, addr, bias) \
> - "ulw %[low32], "#bias"("#addr") \n\t" \
> +#define MMI_LWLRC1(fp, addr, bias, off) \
> + "lwl %[low32], "#bias"+"#off"("#addr") \n\t" \
> + "lwr %[low32], "#bias"("#addr") \n\t" \
> "mtc1 %[low32], "#fp" \n\t"
>
> #define MMI_LWXC1(fp, addr, stride, bias) \
> @@ -68,9 +69,10 @@
> #define MMI_SWC1(fp, addr, bias) \
> "swc1 "#fp", "#bias"("#addr") \n\t"
>
> -#define MMI_USWC1(fp, addr, bias) \
> +#define MMI_SWLRC1(fp, addr, bias, off) \
> "mfc1 %[low32], "#fp" \n\t" \
> - "usw %[low32], "#bias"("#addr") \n\t"
> + "swl %[low32], "#bias"+"#off"("#addr") \n\t" \
> + "swr %[low32], "#bias"("#addr") \n\t"
>
> #define MMI_SWXC1(fp, addr, stride, bias) \
> PTR_ADDU "%[addrt], "#addr", "#stride" \n\t" \
> @@ -79,8 +81,9 @@
> #define MMI_LDC1(fp, addr, bias) \
> "ldc1 "#fp", "#bias"("#addr") \n\t"
>
> -#define MMI_ULDC1(fp, addr, bias) \
> - "uld %[all64], "#bias"("#addr") \n\t" \
> +#define MMI_LDLRC1(fp, addr, bias, off) \
> + "ldl %[all64], "#bias"+"#off"("#addr") \n\t" \
> + "ldr %[all64], "#bias"("#addr") \n\t" \
> "dmtc1 %[all64], "#fp" \n\t"
>
> #define MMI_LDXC1(fp, addr, stride, bias) \
> @@ -90,9 +93,10 @@
> #define MMI_SDC1(fp, addr, bias) \
> "sdc1 "#fp", "#bias"("#addr") \n\t"
>
> -#define MMI_USDC1(fp, addr, bias) \
> +#define MMI_SDLRC1(fp, addr, bias, off) \
> "dmfc1 %[all64], "#fp" \n\t" \
> - "usd %[all64], "#bias"("#addr") \n\t"
> + "sdl %[all64], "#bias"+"#off"("#addr") \n\t" \
> + "sdr %[all64], "#bias"("#addr") \n\t"
>
> #define MMI_SDXC1(fp, addr, stride, bias) \
> PTR_ADDU "%[addrt], "#addr", "#stride" \n\t" \
> @@ -141,17 +145,18 @@
> #define DECLARE_VAR_LOW32 int32_t low32
> #define RESTRICT_ASM_LOW32 [low32]"=&r"(low32),
>
> -#define MMI_ULWC1(fp, addr, bias) \
> - "ulw %[low32], "#bias"("#addr") \n\t" \
> - "mtc1 %[low32], "#fp" \n\t"
> +#define MMI_LWLRC1(fp, addr, bias, off) \
> + "lwl %[low32], "#bias"+"#off"("#addr") \n\t" \
> + "lwr %[low32], "#bias"("#addr") \n\t" \
> + "mtc1 %[low32], "#fp" \n\t"
>
> #else /* _MIPS_SIM != _ABIO32 */
>
> #define DECLARE_VAR_LOW32
> #define RESTRICT_ASM_LOW32
>
> -#define MMI_ULWC1(fp, addr, bias) \
> - "gslwlc1 "#fp", 3+"#bias"("#addr") \n\t" \
> +#define MMI_LWLRC1(fp, addr, bias, off) \
> + "gslwlc1 "#fp", "#off"+"#bias"("#addr") \n\t" \
> "gslwrc1 "#fp", "#bias"("#addr") \n\t"
>
> #endif /* _MIPS_SIM != _ABIO32 */
> @@ -162,8 +167,8 @@
> #define MMI_SWC1(fp, addr, bias) \
> "swc1 "#fp", "#bias"("#addr") \n\t"
>
> -#define MMI_USWC1(fp, addr, bias) \
> - "gsswlc1 "#fp", 3+"#bias"("#addr") \n\t" \
> +#define MMI_SWLRC1(fp, addr, bias, off) \
> + "gsswlc1 "#fp", "#off"+"#bias"("#addr") \n\t" \
> "gsswrc1 "#fp", "#bias"("#addr") \n\t"
>
> #define MMI_SWXC1(fp, addr, stride, bias) \
> @@ -172,8 +177,8 @@
> #define MMI_LDC1(fp, addr, bias) \
> "ldc1 "#fp", "#bias"("#addr") \n\t"
>
> -#define MMI_ULDC1(fp, addr, bias) \
> - "gsldlc1 "#fp", 7+"#bias"("#addr") \n\t" \
> +#define MMI_LDLRC1(fp, addr, bias, off) \
> + "gsldlc1 "#fp", "#off"+"#bias"("#addr") \n\t" \
> "gsldrc1 "#fp", "#bias"("#addr") \n\t"
>
> #define MMI_LDXC1(fp, addr, stride, bias) \
> @@ -182,8 +187,8 @@
> #define MMI_SDC1(fp, addr, bias) \
> "sdc1 "#fp", "#bias"("#addr") \n\t"
>
> -#define MMI_USDC1(fp, addr, bias) \
> - "gssdlc1 "#fp", 7+"#bias"("#addr") \n\t" \
> +#define MMI_SDLRC1(fp, addr, bias, off) \
> + "gssdlc1 "#fp", "#off"+"#bias"("#addr") \n\t" \
> "gssdrc1 "#fp", "#bias"("#addr") \n\t"
>
> #define MMI_SDXC1(fp, addr, stride, bias) \
> @@ -203,6 +208,12 @@
>
> #endif /* HAVE_LOONGSON2 */
>
> +#define MMI_ULWC1(fp, addr, bias) MMI_LWLRC1(fp, addr, bias, 3)
> +#define MMI_USWC1(fp, addr, bias) MMI_SWLRC1(fp, addr, bias, 3)
> +
> +#define MMI_ULDC1(fp, addr, bias) MMI_LDLRC1(fp, addr, bias, 7)
> +#define MMI_USDC1(fp, addr, bias) MMI_SDLRC1(fp, addr, bias, 7)
> +
> /**
> * Backup saved registers
> * We're not using compiler's clobber list as it's not smart enough
> --
> 2.32.0
>
1. Use l/r pair is a good choice.
2. Suggest keep the macro name unchanged, it's easier to use and associated with unaligned access.
3. Another 4 patches in this series are look good to me.</jiaxun.yang at flygoat.com></jiaxun.yang at flygoat.com></jiaxun.yang at flygoat.com>
More information about the ffmpeg-devel
mailing list